ChronoPatch slows vision-language models using only final responses
Researchers describe a black-box attack that makes vision-language models generate longer outputs, raising latency and energy use without needing access to model internals.
In plain words
Vision-language models answer prompts by generating text token by token, so longer answers cost more time and computing. ChronoPatch is an attack that tweaks images so models produce much longer replies, even when an attacker can only see the final response. The authors report this raises serving delays and energy consumption on several models and commercial APIs.
What was announced
The paper, posted to arXiv, argues that because autoregressive decoding makes response length a key driver of latency and compute cost, adversarial visual inputs create an availability risk for generative inference services. Earlier energy-latency attacks, the authors note, typically relied on gradients, logits, or token-level feedback; ChronoPatch instead needs only the final response. It works in two stages: an evolutionary search guided by response length finds visual "delay anchors," then decision-based optimization calibrated in a subspace reduces the size of the perturbation.
In tests reported by the authors, the attack raised inference latency by 47.86% to 293.34% and GPU energy consumption by 41.44% to 249.70% across four open-source vision-language models, using a 5% spatial budget, and produced more repetitive outputs. Against three commercial VLM APIs, it increased response length by 37.6% to 81.5% and client-observed latency by 1.82x to 2.29x, with query budgets of 500 per API. The authors present this as a generation-based availability risk that persists under final-response-only access, and call for generation-aware safeguards in deployed services.
What it does not show
- The evidence is the paper's abstract only; experimental setup, baselines, and peer review are not visible in the supplied excerpt.
- All latency, energy, and response-length figures are author-reported on the tested models and APIs; no independent confirmation is provided.
- The abstract does not name the specific models or commercial APIs tested, nor whether providers have mitigated the attack.
For security and IT teams
- Scope
- The reported findings cover four open-source vision-language models and three commercial VLM APIs, tested under a 5% spatial perturbation budget and query budgets of 500 per API.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.